Abstract

Compositional distributional semantics represents meanings of phrases and sentences by vectors built from representations of the words therein. The categorical framework of Clark, Coecke, and myself offers a method whereby one builds these vectors by transforming the grammatical structure of the phrase/sentence into a linear map. In this framework, meanings of words with functional types become matrices, cubes, and in general higher order tensors. The work of Grefenstette, Kartsaklis, and I showed how different instantiations of the framework improve tasks such as phrase/sentence disambiguation, similarity, and entailment. In recent work with Blundell and Jezek we show how these models are also helpful at the word level by employing them in the verb similarity task of Gerz et al. With Kartsaklis and Ramgoolam, we propose that perturbed Gaussian models with permutation symmetry provide a framework for characterizing the statistical properties of the word tensors. In this talk, we will present the categorical framework and go through snippets of the experiments, with focus on the latter two recent advances.